Appropriate Pixel Size for Orthophotography Hans-Peter Bähr Institute for Photogrammetry and Remote Sensing, University of Karlsruhe Englerstrasse 7, D-7500 Karlsruhe 1, Germany Email: bähr@ipf.bau-verm.uni-karlsruhe.de Things are going to change drastieally by digital image processing. This is beeause 3D-information of the Earth's surfaee is stored in large data bases, no matter what data acquisition system was applied: The data bases provide eompatibility. Images will generally be a very important layer in Geo-Information Systems. Moreover, images are the starting point for automatie data extraction. All these factors eontribute to the fact, that digital orthophotos suddenly are beeoming more and more important for industry and application (COLOMINA et al. 1991, HÖHLE 1992, MAYR 1992). Abstract Pixel size is a basic parameter for digital imagery. It is shown, that there exists no general rule for an appropriate pixel size: It is a function of many parameters, like object frequency, image quality, specific application andeconomical restrictions of date processing and storage. Zusammenfassung Pixelgröße ist ein Basisparameter für digitale Bilder. Es wird gezeigt, daß keine allgemeinen Regeln für geeignete Pixelgrößen existieren: Sie ist eine Funktion vieler Parameter, wie Objektfrequenz, Bildqualität, jeweilige Anwendung und wirtschaftliche Beschränkungen von Bildverarbeitung und speicherung. 1 2 Economical Considerations for Orthophoto Production Reeent presentation of digital orthophoto systems by industry show that production costs apparently have dropped to economy level. This is really prooved by BÄHR/WIESEL 1991, giving approx. US $ 25 computing eost for colour orthophotos of 230 mm x 230 mm standard format plus manpower. The equation Introduction (1) presents the total cost C of an orthophoto as a function of instrument Gi and instrument time ti plus man power cost Gm and operator time t m . Though today the man power component is still higher than the instrument component, there is a chance that manpower costs will drop considerably when automation is more commonly applied. Orthophotos, i. e. rectified photogrammetric imagery, are used sinee many deeades as a substitute or a supplement for topographie maps. There exist without doubt many advantages like quiek and eeonomie production, actual and eomplete information, espeeially for the environmental domain. Considering pixel size of digital orthophotos, it of course enters ti in equation (l). Therefore, i twill always be an important factor, even if instrument cost will continue to go down. Nevertheless, orthophotography was not always generally aceepted like one should have expected (e. g. KELLERSMANN 1985). This was due to eompletely different technology from eonventional eartography: a half-tone paper-print photography is not "compatible" to a eonventional line map. Combinations of both products are offered ("photo maps")and show the way to follow in future: This is a particular factor for digital systems which does not exist for analog systems, where "geometrie resolution" principally does not play an important role for economy, exept for scale considerations, which finally are the same for digital systems. 64 For economical consideration we may approximate the pixel number in a standard photogrammetric image by n [MB] '" (0,5· LP)2, Lp/mm where LP stands for geometrie resolution in linepairs/mm (see BÄHR 1989). This relation is visualised in Fig. 1. Standard values are given by 25 LP /mm and corresponding pixel size of 20 p,m. This quantity may be exactly stored on a 9-track tape of 6 250 bpi. SO Storage will not represent .. a considerable cost factor any more as shown by BAHR/WIESEL 1991, even when producing colour orthophotos. Resampling is a more important factor, infiuencing not only cost but also time, because the operation al environment, time for orthophoto production is most important; "economy" is highly correlated with "time". Adding the working steps to produce an orthophoto digitally, reduction of pixel size from 50 to 25 .. p,m raises the total processing time by 100 % (BAHR/WIESEL 1991). für 10 100 200 300 Pixel (MB) Fig. 1 Data quantity in [MB] as a function of resolution and pixel size for photogrammetric imagery (BÄHR 1989) Therefore, adequate pixel size for orthophotography has to be selected earefully according to the respeetive needs in order to meet eost and time requirements. 3 Pixel size [jJm] 3.1 Theoretical considerations Pixel size is a parameter of image quality, more specifically a parameter of geometrie resolution, whieh is always conneeted to (teehnieal or biologieal) systems. The "real world" has no "resolution", it ean be represented by a fractional model, whieh IS teehnieally infinite. Onee the "real world" is imaged, the imaging proeess is band limited. This means, that in the frequeney domain the resulting image H (u, v) is restricted to a spectral band ±v: Size The term "orthophoto" relates to a weH-known analog reetifieation teehnology, whieh is at its end sinee digital systems have beeome operational. Data type is the eonventional, analog photogrammetric image of 230 mm x 230 mm format. -v ~ H(u, v) ~ +v (2) due to Modulation Transfer whieh in praetise always produces degradation of the image. H(u, v) = F(u, v). G(u, v) The word "orthophotography" defines geometrie properties different from the original central projection of analog photography. An "orthophoto" is a standard product in Photogrammetry, somewhat "historie" on photographie paper. (3) where F (u, v) is the spectrum of an original image, and G (u, v) the transformation funetion. The Modulation Transfer Function MTF = IG(u, v)1 in (3), which designes the geometrie resolution of the system output, is eomposed by the respective MTF's of the subsystems. For eonventional orthophotos we may write for instance: A "digital" orthophoto on the other hand, may be taken as a standard product, too. This is eorreet, when it means only the substitution of analog photographie teehnology by digital eomponents. In this ease, the final product is weB defined, i. e. identieal to the eonventional product "on a sheet of paper" . MTF = MTFa (4) . MT Fgl MT Fe! MT Fed However this is for sure not the eorrect way to handle new ted~nology. Real new teehnology - and digital image processing in photogrammetry is one - will neeessarily lead to new products and provoke new applications, starting may be from the old ones. In the next paragraphs we shall diseuss the appropriate pixel size in the light of ehanging technology and ehanging products. ·MTFol MTFo! MTFod where the indexes stand for "atmosphere" (a), "lens" (I) "film" (f)and "development proeess" (d). 1, fand d appear for both the photogrammetrie eamera (e) and the orthophoto projector (0); the respective fadors are not the same (i. e. MTFeJ t= M,TFoJ). 65 Asking for the appropriate pixel size for the image H (u, v) in (3), it is theoretically given by the NYQUIST frequency v, i. e. the highest frequency transferred by the MTF of the whole imaging process (4). This leads to the well-known Sampling Theorem a< - ..!... 2// 3.2 PracticaI considerations In practise, there is always band limitation as images in any case have passed a technical or physiological system in order to be real. Consequently, we may derive a MTF according to (4); different methods to do this are reported in BÄHR, 1988. (5) giving the descrete pixel values a as a function of the limiting frequency (supposing isotropie character). Fig. 2 explains how to derive the appropriate pixel size from the MTF. After the Sampling Theorem it corresponds to N /2, where N marks the intersection of the MTF with the Resolving Threshold of the Application system (RETAS). The determination of v, being the basic parameter for pixel size, is not trivial. J ust one number to characterize the highly complex data set of a twodimensional image will only present a very rough referenee. This becomes evident, when analysing a straight lines in an image. Lines are the most critical components in (digital) images in theory and in practise, and therefore we will take them for discussing some basic issues. The two curves separate four areas, which caracteristically show the conditions for pixel size: Only in area 1 the information transported by theMTF satisfies the requirement of the RETAS; the areas 2 and 4, below the RETAS curve are principally excluded, i. e. there exist uo pixel size which would correspond to the requirement of the application expressed by the RETAS curve. An "ideal straight line" theoretically contains all frequencies, as it is composed of points without space in between. This may be visualised clearly in the frequency domain, where straight lines in the image appear again as straight lines. Consequently, there is no band limitation for such an "ideal straight line" , and any pixel size would deteriorate H. Area 3 may be varied as a function ofpixel size, which directly designs the MTF. It is very important to point out, that two factors contribute to the appropriate pixel size, the MTF and the RETAS. As far as the MTF is concerned, it relates to the input data, whereas the RETAS relates to the output data. A straight line is a "model from analytical geometry", which means a high level of abstraction. By the way, we run into the same problems for "points", which again are nothing than an abstract model and do not exist in reality. In practise, lines do not correspond rigerously to the ideal model, neither in the real world nor in the "recorded version" on imagery. One should add "fortunately", because otherwise there would be no solution for our question. Having in mind digital orthophoto production, input data may for instance be * Conventional photogrammetric imagery * Photography from amateur camera * Scanner imagery from aerial or space platforrns While scanner imagery is already digital, photography has to be converted into pixels. This process has to take into account not only the MTF of the original photography, but also the RETAS as shown. In other words, the pixel size has to consider the intended further application for the output data, which may for instance be * Produdion of conventional orthophotos, substituting only analog projection by digital resampling * Orthophoto layer in digital data base * Digital orthophotos for separate applications, like environmental monitoring or point measurement o ~------ __________- L___________ N We take the first point for an example: The final result, a paper print orthophoto, has to serve for analysis by a human operator without using visual assistance by lenses etc. The resolution of the human eye affords a pixel size of about 50 ftm for 1,4' viewing angle, given by the theoretical physiological treshold. This relates to a quality which had been guaranteed by the conventional rastering and printing process (see WIESEL, 1985). LP/mm Fig. 2: Pix~l size N/2 as a function of MTF and Resolving Threshold of Application System (RETAS). For the numbers see text. 66 Evidently, the RETAS of this example does not seem to be in accordance with the potential of a photogrammetric image - "its curve in Fig. 2 is too steep" . The human eye is not the adequate sensor for looking at an original photogrammetric image. This of course does change when armed by a lens, i. e. when the hatched area in Fig. 2 grows due to smoother inclination of RETAS. Conditions for pixel size change considerably when the human eye is eliminated from the system. This happens for computer vision, where the input for the grey values after resampling is a digital system. In this case, the signal is practically transformed without any degradation caused by a MTF, and the RETAS does not exist. Consequently, the digital system may use the input data fuHy, without regarding a RETAS. An example is digital correlation of targets (point signals) as shown by BÄHR (1988), where correlation accuracy only depends on the number of pixels which do show gradient effects. This is exclusively a function of scale. 3a 4 Variable Pixel Size in one Scene: The degressive Sampling Approach It is evident, that image scale, geometric resolution or pixel size should be an issue of the respective application. Different from analog photography, digital orthophotography is not necessarily restricted to a uniform resolution in one scene. Within one scene we may vary pixel size due to well-defined demands by operator's request. A very simple example follows from the above mentioned determination of targets in computers vision systems: Here small pixels must be available only in areas elose to the targets. We should remember that the human vision system depicts carefully only a very small portion of the sphere by mecanically moving head and eyeballs and optically focusing the point of interest. 3b The selection of different pixel size within one scene may follow various algorithms, 2 of them have been studied here. It is one of the advantages ofdigital image processing, that the processes are controlled by software and not by hardware. The objective of the following analysis is to visualise the effect from variable pixel size in one scene of a digital orthophoto. The data was scanned from aerial photography of 1 : 6 000 scale at 50 pm, showing the Campus of Karlsruhe University, Germany. Fig. 3 gives the original data for three charaderistic objects: "Cars" (3 a, a parking area and its environment), "building" (3 b, appartment houses) and "forest" (3 c, i. e. part of the gardens elose to the famous Karlsruhe 3c Fig. 3 Test objects from aerial photography 1 : 6 000 50 pm pixel size, area 256 x 256, original (100 %) a: "cars"j b: "building"; c: "forest" 67 4a 5a 4b 5b 4c 5c Fig. 4 Effect by uniform data reduction 2 x 2 patches (25 %) Fig. 5 Effect by uniform data reductioll 4 x 4 patches (6,25 %) 68 6 al 6 a2 6 bl 6 b2 6el 6 c2 Fig. 6 Effect by degressive sampling threshold qlO a 28,4 %, b 15,7 %, C 18,5 % = = = 69 7 al 7 a2 7 bl 7 b2 7cl 7 c2 Fig. 7 Effect by degressive sampling threshold q20 a = 5,9 %, b = 2,8 %, c = 0,7 % 70 Castle). The smaH scenes include 512 x 512 pixels each. Table 1 displays the values obtained for the three scenes, giving r q and rd as weH as different values for t a . In Fig. 4 and 5 the original data have been compressed to patches of 2 x 2 and 4 x 4, conserving 25 % and 6,25 % of the original data, respectively. This uniform approach does not take into account the image content; consequently it shows a general lowpass filtering effect, which very often will not be accepted for orthophotos. For analysing the results, we have 3 references: the images, the respective pixel pattern (Fig. 6 and 7, fight column) and table 1. The visualisation shows that reduction to approx. 20 % of the original pixel size does not lead to considerable degradation. This is not the case when using the uniform reduction (Fig. 4 and 5). Moreover, further reduction to below 10 % is critical as expected. On the other hand, the pixel Different pixel sizes in one scene lead to considerably better results as shown in Fig. 6. The algorithmus used refer to grey value differences. Different grey values of one patch i are substituted by a single one (e. g. the me an value) if pattern show that frequencies of high contrast are still preserved, even for only 1 % of the original pixel quantity. The pixel pattern are very useful for the analysis, as the human eye alone is often mislead. (6) where t a is an arbitrary threshold (in greyvalue units) and ri the result of an algorithm applied to the patch i. Very interesting is the process for the forest scene; here the meadow in the center is substituted by an isolated single pixel al ready for 18,5 % of the information. At this stage (t q = 10); high frequencies of the trees are still preserved. This goes drastically 20 (only 0,7 % of the original down for t q information), where no useful information is any more present. The procedure may start taking the whole scene of 256 by 256 grey values as the first patch. The subsequent patches then are reduced systematicaHy, following a quadtree approach, going from 4 (of 128 x 128 pixels) to finally 16384 (of 2 x 2 pixels). However we have to substract from this number those patches, which have already been assigned to uniform pixels. A simple approach for ri = Fig. 8 demonstrates clearly the different effect for the 3 scenes. High frequencies afford large pixel numbers, and "forest" behaves differently from "man made features" like cars and buildings. would be We call the procedure "degressive sampling", for it starts at smaH pixel sizes and leads to coarser representation for aeras of lower frequency which in general may be less interesting. Anyway the potential of this approach does not end here. Both thresholds t a and algorithms ri may be varied in order to design the required result properly. (7) n where n is the pixel number in one patch; j controls the sum of the individual pixels in that patch and k their difference from the mean value. Pixel Number Instead of the grey value differences their quadratic sum may be taken, too. This leads to the algorithm r q , used for the displayed examples in the figures 6 and 7. The reason for taking r q instead of r d is because of better contrast for printing. This is a consequence from the quadratic term. % so 40 30 20 Quadr. Difference q Linear Difference p Cars Pixel number % Building Pixel number % Forest Pixel number % Original t a = 10 t a = 20 t a = 40 t a = 80 65536 100 18592 28,4 3868 5,9 12313 18,8 2998 4,6 10 4 x 4 atches = 6 25% 20 65536 100 10309 15,7 1867 2,8 8101 12,4 1720 2,6 65536 100 12145 18,5 436 0,7 7372 11,2 709 1,1 40 60 80 Threshold d Fig. 8 Reduction of pixel numbers in % as a function of density differences (threshold d) and uniform pixel clustering (2 x 2 and 4 x 4 patches) Table 1: Pixel quantity in a 256 x 256 scene for different scenes algorithms and thresholds 71 5 Conclusion LITERATURE BÄHR, H.-P.: Measures for Geometrie Resolution of Digital Cameras. ISPRS Congress Kyoto, 1988, CommI Pixel size will always play an important role for orthophoto production, even if some parameters, like data storage, are going to become obsolete. In future, pixel size is not necessarily any more a fixed standard value, but within a scene a flexible function of many parameters, which depend on specific applications. The degressive sampling approach seems to be a flexible procedure which consequently makes use of the potential of digital image processing for orthophoto production, not accepting uniform pixel size for a complete scene. The data reduction effect and the conservation of high frequencies may only be the starting point for a new thinking. The next step will apply itfor image segmentation, as a preprocessing approach. 6 BÄHR, H.-P.: Das digitale Orthophoto und seine Möglichkeiten. Photogrammetric Week, Stuttgart 1989 BÄHR, H.P.and WIESEL, J.: Cost-Benefit Analysis of Digital Orthophotos Technology. In: EBNER/FRITSCH/HEIPKE (Eds.) Digital Photogrammetric Systems, Wichmann Verlag Karlsruhe 1991 WIESEL, J.: Herstellung digitaler Orthophotos. In: Digitale Bildverarbeitung, Karlsruhe 1985 FA UST, H.-W.: Digitalisierung photogrammetrischer Bilder. Zeitschrift für Photogrammetrie und Fernerkundung, 1/1990 S. 6 - 11 HÖHLE, J.: Herstellung von digitalen Orthophotos an einer Arbeitsstation des Geoinformationssystems TIGRIS. Zeitschrift für Photogrammetrie und Fernerkundung S. 42 - 48, 1992 Acknowledgement KELLERSMANN, H.: Farbige Luftbildkarten 1 : 5000 - nein, danke? Erfahrungen über den Verkauf von DrucksWcken. Zeitschrift für Photogrammetrie und Fernerkundung, S. 19 - 22, 1985 The author gratefully thanks Yuval Fisher/San Diego for providing the computer programs for degressive sampling and Heiko J acobs for performing the image processing. MAYR, W.: Das photogrammetrisehe Bildverarbeitungssystem von Carl Zeiss. Zeitschrift für Photogrammetrie und Fernerkundung S. 13 - 18, 1992 COLOMINA, 1.; NAVARRO, J. and TORRE, M.: Digital Photogrammetric Systems at the I. C. C. In: EBNER/FRITSCH/HEIPKE (Eds.): Digital Photogrammetric Systems. Wichmann Verlag Karlsruhe 1991 72